Exploitation of vegetation indices and random forest for cartography of rosemary cover: Application to gourrama region, Morocco

Abstract

The purpose of this study is to give an efficient and practical method for mapping rosemary cover, which belongs to esparto grasslands. The approach consists of merging two technics: remote sensing and machine learning. At first, a set of vegetation indices was calculated using Sentinel 2A MSI image clipped for the study area. In a second place, a correlation test was executed to eliminate highly correlated indices. And so, three combinations were determined using the holden indices. After that, the model was trained using terrain truth samples. In the end, the model was applied to classify the image built, for the study area, by stacking the indices. The model was run for the study area, then validated using terrain sample points. Results are maps of rosemary cover densities corresponding to each combination. The validation test showed a score of 80, 67.5, and 75,5% respectively for combination 1, 2 and 3. This confirmed the selection of the first combination.